Despite the impressive Transaction Processing Performance Council (TPC) scores that Microsoft generated by using scale-out technology and SQL Server's federated-databases feature, it's too soon to count out scale-up technology based on SMP systems. Adding more processing power by scaling up is as simple as buying a bigger, faster system. If you're running a maxed-out 2-way system with dual 450MHz processors, then moving to a brand-new 4-way system running quad 700MHz processors will buy you a lot of headroom. This solution is about as close to a sure bet as you can get in the computer world; it doesn't require redesigning the database, nor does it require any changes to your applications or to your systems management practices. You just back up your current databases and restore them to the new system, and your system will be up and running—faster than ever.
Given the current state of the technology, scaling out isn't nearly as simple as scaling up. You can't just plug in a new cluster node, set the server's Go Fast property to True, and walk away. Implementing a scale-out solution requires manually splitting the database among the cluster nodes—and the partitioning can't be haphazard. First, you need to know your data really well, and that data needs to have some type of natural partition or dividing point that lets you divide it effectively among multiple cluster nodes. After you select the data partition, you must manually move the data to the various cluster nodes. Finally, after you've divvied up the database among the different servers, you need to implement new management practices to effectively back up and monitor each cluster node. Each cluster node is a standalone instance of SQL Server; OLE DB and SQL Server's linked-server capability join the cluster nodes. Although Microsoft has promised to include tools in the next release of SQL Server (code-named Yukon) to automate this process and provide a single management image of the cluster, those tools aren't available today.
Scale-out technology does fill a need at the high end of SMP scalability. Although Windows 2000 Datacenter Server, with its 32-way SMP capabilities, has pushed the scale-up ceiling fairly high, SMP has limits. Scale-out technology pushes beyond this ceiling, and for high-end applications, the added performance is worth the pain required to implement the solution. Scaling out holds promise for the future and can be an effective solution for high-end implementations today, but it's not well suited for everyone. Scaling up is still the simplest and surest way to add capacity to a topped-out system.